| description abstract | Fog imposes adverse effect on driving safety. Traditional visibility measurement methods are expensive and limited to a short distance along the roadway. This study aims to identify visibility levels from foggy road images with deep learning methods. To address the shortage of foggy road image data set, a novel method is proposed to generate synthetic fog images based on point cloud and red, blue, & green (RGB) images. A synthetic foggy roadway image data set, kitti-foggy, containing 10,034 images was created with data from the kitti data set. Performance of the proposed method was compared with the traditional stereo-based method. Three typical image classification convolutional neural networks, including ResNet34, ResNet101, and Inception V4, were used to train the data set, and several evaluation matrices were used to evaluate their performances. The proposed method outputs more natural and authentic fog images. ResNet34 demonstrated the best performance among three algorithms with an overall accuracy of about 93%. Real data from a driving recorder and drones was used to verify the capability of ResNet34 to detect real fog. Findings of this study assist in the field of autonomous driving as well as intelligent transportation. | |
